Problem

The increasing success of Artificial Intelligence (AI) systems, particularly in classification tasks, is often hampered by a significant challenge: the lack of explainability. This limitation is particularly critical in high-stakes fields such as healthcare and finance, where understanding the rationale behind an AI's decision is paramount. Many deep learning models, including Classical Boltzmann Machines (CBMs), struggle to provide transparent insights into their decision-making processes, creating a barrier to trust and adoption. This research aims to address this fundamental problem by developing an AI framework that not only performs well but also offers substantive transparency regarding its internal workings.

Method

The proposed solution involves a novel explainable AI framework that integrates principles of quantum computing with classical machine learning, primarily through a comparative study of Quantum Boltzmann Machines (QBMs) and Classical Boltzmann Machines (CBMs). The methodology follows a four-step pipeline: First, the input data, consisting of two digits (0 and 1) from the MNIST dataset, was binarized and subjected to Principal Component Analysis (PCA) to reduce its dimensionality to four principal components. Second, a quantum hidden layer was designed using PennyLane to construct a quantum circuit with strongly entangling layers. This circuit utilizes angle embedding to map classical input features into quantum states, enabling the capture of complex feature interactions through superposition and entanglement. Third, a hybrid Quantum Boltzmann Machine (QBM) architecture was constructed by combining classical layers with this quantum layer, which generates latent features. Finally, for interpreting feature importance, QBMs employed gradient-based saliency maps, which calculate gradients of the model output with respect to the input to reveal influential features. For CBMs, SHAP (SHapley Additive exPlanations) values were used as a classical baseline, although the paper notes its limitations for models with tightly interlinked nodes. To quantify the focus of each model on input features, the entropy of the normalized feature importance values was computed.

Results

The comparative study between QBMs and CBMs yielded distinct results in both classification accuracy and interpretability. QBMs demonstrated significantly higher classification accuracy at 83.5% compared to CBMs, which achieved 54.0% accuracy on the binarized, PCA-reduced MNIST dataset. This substantial performance boost suggests that QBMs are more capable of learning complex patterns within reduced feature spaces, likely due to their quantum-enhanced representational power. Furthermore, in terms of interpretability, QBMs exhibited more concentrated distributions in feature attributions, as evidenced by a lower entropy value of 1.2704 compared to CBMs' entropy of 1.3820. This indicates that QBMs focused their attention on fewer, more relevant "active ingredient" features (specifically PC0 and PC2), providing clearer identification of the most important features driving decisions. A t-SNE visualization of the QBM's hidden states also showed clear and separable clustering for the two digit categories, confirming the model's ability to learn class-discriminative quantum embeddings.

Implications

The research highlights that quantum-classical hybrid models like Quantum Boltzmann Machines can simultaneously achieve improvements in both predictive accuracy and interpretability, addressing a critical need in the field of Explainable AI. The ability of QBMs to provide clearer and more concentrated insights into the "active ingredients" that drive model decisions contributes significantly to the development of more trustworthy and transparent AI systems. This work suggests that quantum-enhanced models offer an intrinsically different and potentially more capable approach to model interpretability compared to traditional classical methods. Future research directions include applying this framework to increasingly complex datasets, refining quantum circuit topologies, and exploring the integration of both classical and quantum interpretability techniques into more comprehensive hybrid XAI systems.